基于知识迁移的数据驱动迭代学习模型预测控制

Data-driven iterative learning model predictive control based on knowledge transfer

迭代学习模型预测控制(iterative learning model predictive control,ILMPC)作为一种广泛应用于批次生产过程的数据驱动智能控制策略,能够在沿批次逐步提高跟踪性能的同时沿时间不断抑制实时干扰.现有ILMPC算法的点对点学习机制依赖于批次运行工况的强一致性,以此保证当前批次与历史批次间的有效信息传递.然而,生产需求和生产环境的变化通常会导致各批次的操作轨迹和操作周期存在差异,从而使得历史批次提供的先验知识对于后续批次呈现出不精确性和不完整性.为了提高ILMPC在变运行工况条件下的适应性和灵活性,本文提出了一种具有知识迁移机制的数据驱动ILMPC策略.建立自适应深度神经网络(deep neural network,DNN)沿批次学习ILMPC控制行为,实现历史控制经验在当前批次工况下的全面转换.为抑制DNN前期估计误差的影响,在知识迁移机制下进一步构建Tube控制结构下的ILMPC算法,保证ILMPC系统的时域稳定性和迭代域收敛性.针对非线性注塑过程的仿真实验验证了在操作轨迹和操作周期同时变化时,所提方法在跟踪精度和收敛速度方面具有明显优势.

Iterative learning model predictive control(ILMPC)has become an excellent data-driven intelligent control strategy for batch manufacturing,featured by progressively improving tracking performance along trials while persistently rejecting stochastic disturbance along time.The point-to-point learning mechanism of the existing ILMPCs generally relies on the strict identity of operating conditions along trials to guarantee the integrity and accuracy of historical data.However,the variation of production requirements and environments usually leads to trial-varying reference trajectories and operating durations,resulting in incomplete and inaccurate historical information for the iterative learning of subsequent trials.To promote the adaptability and flexibility of ILMPC under unconformable prior information,this paper proposes an innovative data-driven ILMPC based on knowledge transfer.The control actions of ILMPC are imitated along trials by an adaptive deep neural network(DNN),based on which the prior operation data are transformed to accommodate the condition of each current trial.Under this knowledge transfer mechanism,the tube control scheme is further integrated into ILMPC to restrain the influence of the considerable DNN approximation error in the early trials,which ensures the timedomain stability and iteration-domain convergence of the ILMPC system.Simulations on the nonlinear injection molding process verify that the proposed method has noticeable advantages in the aspects of tracking accuracy and convergence rate when faced with significant changes of operating reference and duration.